Ant's Ling-3.1-Flash: 560B MoE tops the open field, free trial live
Ant Group's hybrid-reasoning MoE runs 25B active params and edges GPT-5.6-sol on Terminal-Bench 4.0. Free via Vercel now; weights open-source after the trial.

Copy markdown
A 560B MoE with 25B active and 1M context
Ant Group's InclusionAI unveiled Ling-3.1-flash, a ~560B-parameter mixture-of-experts that fires only ~25B params per token and handles up to 1M tokens of context (256K during the launch trial). It's a hybrid-reasoning model built for coding, multi-step analysis, and tool-using agents.
It edges GPT-5.6-sol on Terminal-Bench 4.0
On agent and coding evals, Ling-3.1-flash reportedly outscores DeepSeek-V4.1-Flash, GLM-5.3-Flash and Kimi K3 — and even proprietary GPT-5.6-sol — on Terminal-Bench 4.0, with 75.16 on FrontierSWE and 1,673 Elo on GDPVal-AA v2.1. Numbers are vendor-reported for now, but they put a soon-to-be-open model in frontier coding territory.
Free to call today on Vercel AI Gateway
You can hit it right now through the Vercel AI Gateway (via Novita) or Novita's own API — free during the launch promo running through Oct 13, at ~91 tokens/sec and ~2.2s latency, up to 32K output tokens. Grab a key from the Vercel dashboard and drop it into the AI SDK without changing your integration.
Open weights are 'soon,' not today
The catch for self-hosters: InclusionAI says it'll open-source the model and unlock the full 1M window only after the two-week trial. Until then it's API-only, so if your weekend project needs local weights, keep a fallback ready and watch for the Hugging Face drop.